Aquisição Progressiva de Habilidades por meio de Curriculum Learning para Futebol de Robôs Multiagente

Abstract

This work investigates the integration of Curriculum Learning with Self-play for rein- forcement learning in the context of SSL-EL robot soccer. The research addresses the challenge of developing efficient policies in complex multi-agent environments by proposing a structured methodology that decomposes learning into progressive stages. The implemented framework establishes adaptive criteria for transitioning between tasks, allowing agents to initially develop fundamental skills before facing complete competitive scenarios. The experimental results clearly demonstrate the superiority of the combined approach, with significantly higher win rates in competitive tournaments compared to traditional Full Self-play, as well as an expressive increase in the average goals per match. Additionally, a substantial reduction in total training time and greater stability in the learning process were observed, evidenced by metrics such as policy entropy, policy loss, and explained variance. The analyses confirm that Curriculum Learning provides a solid technical foundation that enhances the benefits of Self-play, resulting in agents with more sophisticated and efficient tactical capabilities.